A meniscus injury prediction method and system based on a neural network

Abstract

The application discloses a kind of meniscus injury prediction method and system based on neural network, the method includes obtaining knee MRI image, based on convolutional neural network, the knee MRI image is automatically segmented, based on the image after segmentation, extract multiple image features, based on SNN network model, construct meniscus injury prediction model and carry out model training and verification.The application realizes to improve meniscus injury diagnostic accuracy, diagnostic efficiency, realizes future injury risk prediction, provides valuable prediction information for clinical doctors, helps to formulate prevention and treatment scheme, simultaneously introduces imageomics features and weight-bearing structure feature heat map and mixed attention mechanism, enhances clinical interpretability, realizes full-automatic diagnosis process.

Description

Technical Field

[0001] The present invention relates to the field of medical image processing, and in particular to a meniscus injury prediction method based on a neural network. Background Art

[0002] Meniscus injury is a common knee injury, especially among those who engage in high-intensity exercise. Traditionally, the diagnosis of meniscus injury relies primarily on clinician experience and manual analysis of knee MRI images. This method is not only highly subjective, but also inefficient and has low diagnostic accuracy.

[0003] Existing technologies mainly focus on three-dimensional reconstruction, segmentation and positioning of meniscus injuries, as well as detection and diagnosis of current injuries. However, there are still deficiencies in predicting future injury risks and further improving diagnostic efficiency and accuracy. Summary of the Invention

[0004] (1) Technical problems solved

[0005] In order to solve the above technical problems, the present invention provides a meniscus injury prediction method based on a neural network.

[0006] (2) Technical solution

[0007] In order to solve the above-mentioned technical problems and achieve the purpose of the invention, the present invention is implemented through the following technical solutions:

[0008] A method for predicting meniscus injury based on a neural network comprises the following steps:

[0009] S1: Acquire knee MRI images. Use a standardized MRI scanning protocol to acquire MRI images of the patient's knee.

[0010] S2: automatic image segmentation, automatically segmenting the knee MRI image based on a convolutional neural network to identify and separate the weight-bearing structures of the knee joint;

[0011] S3: Radiomics feature extraction, extracting multiple image features based on the segmented image, including shape features, texture features, and image intensity features;

[0012] S4: Construct an injury prediction model. Construct a meniscus injury prediction model based on the SNN network model.

[0013] S5: Prediction model training and validation: The model is trained using the knee MRI image dataset, and the model is validated and optimized using internal and external datasets.

[0014] Furthermore, the step S2 further includes:

[0015] S21: Automatic MRI segmentation based on 2D and 3D convolutional neural networks;

[0016] Specifically, it includes preliminary segmentation and secondary segmentation. The preliminary segmentation segments the MRI two-dimensional image through a two-dimensional convolutional neural network, and the secondary segmentation performs fine segmentation of the local area based on a three-dimensional convolutional neural network to segment the knee weight-bearing tissue.

[0017] S22: Adjust the segmentation results generated by the convolutional neural network based on the statistical shape model;

[0018] S23: The automatically segmented MRI image enters the ROI region definition module, and the accuracy of the knee weight-bearing structure segmentation is determined based on the Dice coefficient according to the visual imaging.

[0019] Furthermore, the step S3 further includes:

[0020] The shape features include area, volume, surface area, and aspect ratio; the texture features include grayscale change, contrast, and roughness; and the image intensity features include mean, median, standard deviation, skewness, and kurtosis.

[0021] Furthermore, the step S4 also includes screening out features that are significantly correlated with meniscus injury based on the minimum absolute value convergence and selection operator algorithm.

[0022] Furthermore, it also includes using a selection operator algorithm to perform preliminary feature screening, removing features that contribute little to the prediction or are redundant, and performing multimodal feature fusion on features from different sources to generate a unified feature vector.

[0023] Furthermore, the construction of the meniscus injury prediction model based on the SNN network model also includes: using the BFGS optimization algorithm to update the network weights.

[0024] Furthermore, the meniscus injury prediction model outputs current injury diagnosis and future injury risk assessment.

[0025] Furthermore, step S5 also includes performing a prediction accuracy evaluation based on an AUC evaluation model, and the input content of the AUC evaluation model includes radiomic features, clinical features and time dimension features; the radiomic features include radiomic features of the femur, femoral cartilage, tibia, tibial cartilage, medial meniscus and lateral meniscus, namely shape features, texture features and strength features; clinical features include age, gender, body mass index, history of trauma, history of meniscus injury and history of knee surgery; and time dimension features include baseline MRI imaging data and later follow-up data.

[0026] The present invention also provides a meniscus injury prediction system based on a neural network, which includes: an image acquisition module for acquiring a knee MRI image;

[0027] An image segmentation module for automatically segmenting the knee MRI image based on a convolutional neural network to identify and separate the weight-bearing structures of the knee joint; including automatically segmenting the MRI based on two-dimensional and three-dimensional convolutional neural networks, adjusting the segmentation results generated by the convolutional neural network based on a statistical shape model, and determining the segmentation accuracy;

[0028] Radiomics feature extraction module, which is used to extract multiple image features based on the segmented image, including shape features, texture features, and image intensity features;

[0029] The injury prediction model construction module is used to screen out features that are significantly correlated with meniscus injury based on the minimum absolute value convergence and selection operator algorithm; and to construct a meniscus injury prediction model based on the SNN network model;

[0030] The prediction model training and validation module is used to train the model using the knee MRI image dataset and to validate and optimize the model using internal and external datasets.

[0031] In addition, to achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, on which program instructions of a meniscus injury prediction method based on a neural network are stored. The program instructions of the meniscus injury prediction method based on a neural network can be executed by one or more processors to implement the steps of the meniscus injury prediction method based on a neural network as described above.

[0032] (3) Beneficial effects

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] (1) Improve diagnostic accuracy: By using convolutional neural networks to automatically segment knee MRI images, identify the weight-bearing structures of the knee joint, extract key imaging features, and combine them with shallow neural network algorithms to build a high-precision meniscus injury prediction model to improve the accuracy of diagnostic results.

[0035] (2) Improve diagnostic efficiency: Use fully automated image processing and analysis methods to reduce the workload of doctors, shorten diagnosis time, achieve rapid and efficient diagnosis of meniscus injuries, and improve clinical work efficiency.

[0036] (3) Predict future injury risks: Through long-term monitoring and analysis of patients' knee MRI images, neural network algorithms are used to predict the risk of future meniscus injuries, providing clinicians with valuable predictive information to help develop prevention and treatment plans.

[0037] (4) Enhance clinical interpretability: By introducing radiomics features, weighted structural feature heat maps, and hybrid attention mechanisms into the model, the interpretability of diagnostic decisions is provided, enabling doctors and patients to understand the decision-making process of the model and improving the acceptance and trust of the model in actual clinical applications.

[0038] (5) Realize a fully automated diagnostic process: Integrate image acquisition, automatic segmentation, feature extraction, model training and prediction to build an integrated meniscus injury prediction system, realize a fully automated process from image acquisition to diagnostic result output, and enhance the practicality and promotion of the system. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0040] Figure 1 This is a flowchart of a method for predicting meniscus injury based on a neural network according to an embodiment of the present application;

[0041] Figure 2 This is a diagram of a deep learning automatic segmentation scheme for the weight-bearing structure of the knee using nuclear magnetic resonance according to an embodiment of the present application;

[0042] Figure 3 is a heat map of characteristics of a single load-bearing structure according to an embodiment of the present application;

[0043] Figure 4 It is a complete weight-bearing structure characteristic heat map and a meniscus injury performance map predicted by the final model according to the embodiment of the present application;

[0044] Figure 5a 、 5b 5c is a flowchart of the operation of the final model-assisted diagnosis system according to an embodiment of the present application and the improvement of the surgeon's ability to predict meniscus injury with the assistance of the final model. DETAILED DESCRIPTION

[0045] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0046] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that, in the absence of conflict, the following embodiments and features in the embodiments can be combined with each other. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present disclosure.

[0047] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present disclosure. The illustrations only show components related to the present disclosure and are not drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component can be changed at will, and the component layout type may also be more complicated.

[0048] See also Figure 1 ,A meniscus injury prediction method based on a neural network comprises the following steps:

[0049] S1: Acquire knee MRI images; Acquire knee MRI images of patients using standardized MRI scanning protocols to ensure image quality and consistency, providing a reliable basis for subsequent image processing and analysis.

[0050] Optional, based on MRI scanners, SAG-3D-DESS-WE and PDWI sequences are acquired through standard protocols to ensure image quality and consistency.

[0051] S2: Automatic image segmentation:

[0052] Automatically segment the knee MRI image based on a convolutional neural network to identify and separate the weight-bearing structures of the knee joint. This specifically includes the following steps:

[0053] S21: Automatic MRI segmentation based on 2D and 3D convolutional neural networks;

[0054] Specifically, it includes preliminary segmentation and secondary segmentation. The preliminary segmentation segments the MRI two-dimensional image through a two-dimensional convolutional neural network, and the secondary segmentation performs fine segmentation of the local area based on a three-dimensional convolutional neural network to segment the knee weight-bearing tissue, thereby improving the accuracy of automatic segmentation.

[0055] The local area includes the contact area between bone and cartilage, and cartilage and meniscus, and the knee weight-bearing tissue includes the femur, femoral cartilage, tibia, tibial cartilage, medial meniscus and lateral meniscus.

[0056] S22: Adjust the segmentation results generated by the convolutional neural network based on the statistical shape model;

[0057] Specifically, it involves capturing shape variations between different individuals by learning anatomical data of the knee joint structure.

[0058] Recognized gold-standard manually segmented datasets were used as training sets, including the Imorphics gold standard (IGS) dataset (88 medial and lateral meniscus MRI segmentation images using SAG-3D-DESS-WE and PDWI sequences), the Segmentation of Knee Images 2010 (SKI10) dataset (150 femur, femoral cartilage, tibia, and tibial cartilage segmentation images using T1, T2, gradient echo, and spoiled-gradient echo sequences), and the OAI-Zuse Institute Berlin (ZIB) dataset (507 femur, femoral cartilage, tibia, and tibial cartilage segmentation images using SAG-3D-DESS-WE sequences). Unsegmented images served as validation sets. Automatically segmented images were generated using a convolutional neural network (CNN) algorithm. The CNN algorithm employed a sequence of 2D convolutional neural networks (2D CNNs), statistical shape model (SSM) adjustments, 3D convolutional neural networks (3D CNNs), and SSM post-processing to generate prior knowledge.

[0059] The above method ensures the anatomical rationality and consistency of the segmentation results and improves the accuracy of the segmentation results. Especially in the presence of noise, artifacts or unclear boundaries, the statistical shape model can help ensure the anatomical rationality of the segmented shape.

[0060] S23: The automatically segmented MRI image enters the ROI region definition module, and the accuracy of the knee weight-bearing structure segmentation is determined based on the visual imaging.

[0061] The accuracy of automatic segmentation of the knee weight-bearing structures was determined by comparing the similarity between the automatic segmentation and the manually segmented gold standard dataset using the Dice coefficient, where 0 represents complete dissimilarity and 100% represents complete similarity. The Dice coefficient is calculated as: Dice(s1,s2) = 2*comm(s1,s2) / (leng(s1)+leng(s2)). Where comm(s1,s2) is the number of identical characters in s1 and s2; leng(s1) and leng(s2) are the lengths of the strings s1 and s2. A value above 80% is considered high for automatic segmentation. Currently, the Dice coefficient for this model is >83.0%.

[0062] like Figure 2 This is a diagram of the deep learning automatic segmentation solution for the knee weight-bearing structure by nuclear magnetic resonance (NMR) according to an embodiment of the present invention. Figure 2 From left to right, the segmentation schemes are the femur, femoral cartilage, tibia, tibial cartilage, medial meniscus, and lateral meniscus. Figure 2 From top to bottom are the coronal images of the knee MRI and the three-dimensional shape after image segmentation.

[0063] like Figure 3 This example shows a heatmap of knee MRI pixel values ​​for a single weight-bearing structure. This heatmap allows model users to more directly identify areas of meniscus damage, providing important interpretability for the model. The heatmap uses Python to extract the MRI image of the ROI region and import it into 3D-Slicer software, using volume rendering to create 2D and 3D heatmaps of the weight-bearing structure.

[0064] This method achieves precise automatic segmentation of the weight-bearing structures of the knee joint by combining multi-level segmentation with two-dimensional and three-dimensional convolutional neural networks, the introduction of statistical shape models, and a multi-step post-processing mechanism. This reduces the subjectivity and errors inherent in manual operation and improves segmentation accuracy and efficiency. Furthermore, a hybrid attention mechanism is introduced into the model to generate a heat map of weight-bearing structure features, which visually captures attention maps. This enhances the model's clinical interpretability and improves its acceptance and trust in practical applications.

[0065] S3. Radiomics Feature Extraction:

[0066] Extract multiple image features based on the segmented image, including shape features, texture features, and image intensity features. The specific steps include:

[0067] S31: image preprocessing, including normalization and filtering;

[0068] Preprocessing operations are used to reduce intensity differences in the image and standardize the size and orientation of the image;

[0069] S32: Specific regions of interest (ROIs) were extracted, including the femur, femoral cartilage, tibia, tibial cartilage, medial and lateral menisci;

[0070] S33: Extracting shape features of the image, including area, volume, surface area, and aspect ratio;

[0071] S34: Extracting texture features of the image; specifically, extracting texture features of the image based on a gray-level co-occurrence matrix or a local binary pattern, and quantifying grayscale changes, contrast, and roughness of the image; texture features are used to identify pathological changes in tissue.

[0072] S35: extracting image intensity features;

[0073] Based on the intensity values ​​of pixels or voxels in the image, intensity distribution features are extracted, including mean, median, standard deviation, skewness, and kurtosis. Intensity features can reflect tissue density or signal changes.

[0074] The above steps comprehensively extract multiple features of knee joint images, provide rich input data for subsequent model training and prediction, and improve feature expression capabilities.

[0075] S4. Construct a damage prediction model, comprising the following steps:

[0076] S41: Feature screening;

[0077] This includes screening out features that are significantly correlated with meniscus injury based on the Least Absolute Value Convergence and Selection Operator (LASSO) algorithm;

[0078] In MRI feature extraction, multi-level and multi-structural feature extraction is used, including morphological features such as the volume, surface area, and aspect ratio of the femur, tibia, femoral cartilage, tibial cartilage, and meniscus. In texture feature extraction, gray-level co-occurrence matrix (GLCM) and local binary pattern (LBP) are used to extract MRI texture information of the femur, tibia, femoral cartilage, tibial cartilage, and meniscus tissues. Statistical features such as mean, standard deviation, and skewness are extracted from the MRI pixel value distribution.

[0079] In addition, in terms of feature screening and dimensionality reduction, this model uses LASSO regression for preliminary feature screening to remove features that contribute little to prediction or are redundant. In addition, in terms of feature fusion, features from different sources (morphology, texture, intensity, clinical data) are fused into multimodal features to generate a unified feature vector.

[0080] Traditional radiomics analysis often uses PCA (principal component analysis) or simple Least Absolute Shrinkage and Selection Operator (LASSO) regression for feature dimensionality reduction. However, these methods may overlook key features, leading to information loss. Furthermore, some models directly use raw feature vectors without effective filtering, resulting in feature redundancy and impacting model performance. The advantage of the feature screening scheme presented in this paper is that it improves the validity of feature vectors, reduces interference from irrelevant or redundant features, optimizes feature expression, and ensures that the model captures key features highly relevant to meniscus injury. Multimodal fusion enhances the comprehensiveness and robustness of features.

[0081] S42: Construct a meniscus injury prediction model based on the SNN network model. The model inputs radiomics features and outputs prediction results, including current injury diagnosis and future injury risk assessment. The SNN network model contains 1 hidden layer and 1 output layer. The hidden layer contains 50 neurons, avoiding a complex hierarchical architecture. The activation function uses Tanh to ensure that the network can learn the nonlinear relationship of the data. In addition, the BFGS (Broyden-Fletcher-Goldfarb-Shanno) optimization algorithm is used to update the network weights to reduce the loss function value. The output layer is a Sigmoid activation unit, which outputs a probability value between 0 and 1, indicating the probability of meniscus injury in the next four years. The L2 regularization method is used to control overfitting to prevent the network from overfitting on small sample data sets.

[0082] BFGS, a quasi-Newton method for optimizing the weights and biases of neural networks, is well-suited for small sample datasets, exhibits rapid convergence, and requires relatively few parameter adjustments. BFGS performs particularly well when the neural network has few layers and the feature vectors have been well preprocessed. This application utilizes a lightweight shallow neural network (SNN). Most existing prediction models utilize deep neural networks (DNNs). While DNNs can learn complex nonlinear mappings, they are prone to overfitting in small sample scenarios. Furthermore, some studies have used traditional machine learning models (such as logistic regression or support vector machines), which are unable to effectively capture the complex associations within MRI imaging features. This model has fewer parameters and low computational complexity, is suitable for small and medium-sized data sets, and has a short training time and can run efficiently on ordinary hardware; the SNN algorithm is suitable for tasks where feature engineering has been completed and features are clear, performs well on small sample data sets, and is particularly suitable for medical imaging feature prediction tasks; because the network is shallow and has fewer parameters, the model is less likely to overfit, and is more suitable for medical imaging tasks with limited sample sizes; in addition, the SNN algorithm weight matrix is ​​relatively intuitive, the contribution of each feature can be quantified and explained, and the model decision path is clearer, which is easier for clinicians to understand and verify; finally, SNN has good generalization ability on small and medium-sized data sets, can quickly adapt to new data, and reduce adjustment costs; in addition, L2 regularization is introduced to effectively prevent overfitting. The present invention improves the generalization ability of the model by lightweight SNN, reduces the overfitting problem on small sample data sets, and compared with DNN, the model structure provides higher interpretability and is easy to track the contribution of features in the prediction results.

[0083] S5. Prediction model training and validation:

[0084] A large-scale knee MRI image dataset was used for model training, and the model was validated and optimized using internal and external datasets to ensure its accuracy and robustness.

[0085] Furthermore, it also includes model performance evaluation. The prediction model evaluates the prediction accuracy of the model through the area under the receiver operating characteristic curve (AUC). Specifically, in a data set with M positive samples and N negative samples, there are a total of M*N pairs of samples (a pair of samples is a positive sample and a negative sample). The number of the predicted probabilities of the positive samples in these M*N pairs of samples is counted and calculated according to the formula , in represents the ranking number of the i-th positive sample; M*(1+M) / 2 is the expected total ranking of positive samples. The closer the AUC value is to 1, the stronger the predictive ability of the model. When the AUC value is 0.5, the predictive ability of the model is the same as random guessing.

[0086] The inputs to the AUC evaluation model include radiomic features, clinical features, and temporal features. The radiomic features include shape, texture, and strength features of the femur, femoral cartilage, tibia, tibial cartilage, medial meniscus, and lateral meniscus. Clinical features include age, gender, body mass index, history of trauma, meniscus injury, and knee surgery. Temporal features include baseline MRI data and follow-up data. Temporal data are crucial for building long-term prediction models, as they can capture the progression of meniscus injury in patients.

[0087] Traditional model training methods are relatively simple and fail to fully utilize data from small sample sizes. This paper utilizes K-Fold Cross Validation (K-Fold Cross Validation) to validate the model: the training data is divided K times to ensure consistent model performance across different data subsets, and an internal validation set is used to evaluate the model's training effectiveness and generalization capabilities. The advantages of this model training method lie in its optimized training strategy, which improves model training stability and convergence speed, and its multi-layer validation improves model reliability and generalization.

[0088] When it comes to model evaluation and validation, common metrics like accuracy may not fully reflect model performance. This invention generates a heat map of meniscus MRI characteristics, visualizing the model's predictions and helping doctors intuitively understand the underlying basis for the model's predictions. Combined with heat map visualization, this enhances model interpretability, helping doctors make precise clinical interventions.

[0089] like Figure 4 Shown are the heat maps of the complete weight-bearing structure characteristics of the present invention and the final model predicting the meniscus injury performance. Figure 4 af is the complete weight-bearing structure feature heat map, through Figure 3 The results are superimposed to form a 2D and 3D feature heat map of the femur, femoral cartilage, tibia, tibial cartilage, lateral meniscus and medial meniscus. Figure 4Figure g shows the AUC curve of the final model in the test cohort, which included 84 cases (34 cases in the case group and 50 cases in the control group). The AUC value of the final model was 0.914. Panel a shows the definition of the ROI region on the case group knee MRI, panel b shows the 2D feature heatmap of the intact weight-bearing structure on the case group knee MRI, and panel c shows the 3D feature heatmap of the intact weight-bearing structure on the case group knee MRI. Panel d shows the definition of the ROI region on the case group knee MRI, panel e shows the 2D feature heatmap of the intact weight-bearing structure on the case group knee MRI, and panel f shows the 3D feature heatmap of the intact weight-bearing structure on the case group knee MRI. Panel g shows the AUC of the final model in the test cohort.

[0090] Figure 5 is a flowchart of the operation of the final model-assisted diagnosis system of an embodiment of the present invention and the improvement of the surgeon's ability to predict meniscus injury with the assistance of the final model. Inset a shows the workflow of the final model-assisted system. From left to right, the patient obtains a knee MRI. The orthopedic surgeon can use the final model-assisted system to predict meniscus injury for the patient, or he can use the final model-assisted system without the help of the final model-assisted system. In addition, the final model-assisted system includes a final model and a hybrid attention mechanism based on feature heat maps of weight-bearing structures. Inset b shows the average sensitivity and average specificity of orthopedic surgeons in predicting meniscus injury in patients after using the final model-assisted system in the test cohort. Inset c shows the sensitivity and specificity of individual orthopedic surgeons in predicting meniscus injury in patients after using the final model-assisted system in the test cohort. The results in Figure c show that before the assistance of the final model (AUC of 0.914), the average sensitivity of the seven orthopedic surgeons in predicting that patients would suffer from meniscus injury within four years was 41.2%, and the average specificity was 49.9%. After the assistance of the final model, the average sensitivity of the seven orthopedic surgeons in predicting that patients would suffer from meniscus injury within four years increased to 78.2%, and the average specificity increased to 71.0%.

[0091] In this embodiment, an efficient, accurate, and fully automated meniscus injury prediction method is provided, which significantly improves the accuracy and efficiency of diagnosis, and has significant beneficial effects in terms of predicting future injury risks, enhancing clinical interpretability, realizing fully automated diagnostic processes, reducing medical costs, and improving patient satisfaction, providing strong technical support and service guarantees for clinicians and patients.

[0092] The embodiment of the present invention further provides a meniscus injury prediction system based on a neural network, comprising:

[0093] An image acquisition module, which is used to acquire knee MRI images;

[0094] An image segmentation module for automatically segmenting the knee MRI image based on a convolutional neural network to identify and separate the weight-bearing structures of the knee joint; including automatically segmenting the MRI based on two-dimensional and three-dimensional convolutional neural networks, adjusting the segmentation results generated by the convolutional neural network based on a statistical shape model, and determining the segmentation accuracy;

[0095] Radiomics feature extraction module, which is used to extract multiple image features based on the segmented image, including shape features, texture features, and image intensity features;

[0096] The injury prediction model construction module is used to screen out features that are significantly correlated with meniscus injury based on the minimum absolute value convergence and selection operator algorithm; and to construct a meniscus injury prediction model based on the SNN network model;

[0097] The prediction model training and validation module is used to train the model using the knee MRI image dataset and to validate and optimize the model using internal and external datasets.

[0098] In addition, an embodiment of the present invention also proposes a computer-readable storage medium, on which program instructions of a meniscus injury prediction method based on a neural network are stored. The program instructions of the meniscus injury prediction method based on a neural network can be executed by one or more processors to implement the steps of the meniscus injury prediction method based on a neural network as described above.

[0099] The embodiments described above are merely descriptions of preferred implementations of the present invention and are not intended to limit the scope of the present invention. Without departing from the spirit of the present invention, various modifications and improvements made to the technical solutions of the present invention by ordinary technicians in this field should fall within the scope of protection determined by the claims of the present invention.

Claims

1. A meniscus injury prediction method based on neural network, characterized in that: The steps include: S1: Acquire knee MRI images. Use a standardized MRI scanning protocol to acquire MRI images of the patient's knee. S2: automatic image segmentation, automatically segmenting the knee MRI image based on a convolutional neural network to identify and separate the weight-bearing structures of the knee joint; The step S2 further includes: S21: Automatic MRI segmentation based on 2D and 3D convolutional neural networks; Specifically, it includes preliminary segmentation and secondary segmentation. The preliminary segmentation segments the MRI two-dimensional image through a two-dimensional convolutional neural network, and the secondary segmentation performs fine segmentation of the local area based on a three-dimensional convolutional neural network to segment the knee weight-bearing tissue. S22: Adjust the segmentation results generated by the convolutional neural network based on the statistical shape model; S23: The automatically segmented MRI image enters the ROI region definition module, which introduces a hybrid attention mechanism to generate a heat map of the weight-bearing structure features. Based on the visual imaging and the Dice coefficient, the accuracy of the knee weight-bearing structure segmentation is determined. S3: Radiomics feature extraction, extracting multiple image features based on the segmented image, including shape features, texture features, and image intensity features; S4: Construct an injury prediction model based on a lightweight shallow neural network model to construct a meniscus injury prediction model; S5: Prediction model training and validation: The model is trained using the knee MRI image dataset, and the model is validated and optimized using internal and external datasets.

2. The method for predicting meniscus injury based on neural network according to claim 1, characterized in that: The step S3 further comprises: The shape features include area, volume, surface area, and aspect ratio; the texture features include grayscale change, contrast, and roughness; and the image intensity features include mean, median, standard deviation, skewness, and kurtosis.

3. The method for predicting meniscus injury based on neural network according to claim 1, characterized in that: The step S4 further includes screening out features that are significantly correlated with meniscus injury based on the minimum absolute value convergence and selection operator algorithm.

4. The method for predicting meniscus injury based on neural network according to claim 3, characterized in that: It also includes the use of selection operator algorithms for preliminary feature screening, removing features that contribute little to prediction or are redundant, and performing multimodal feature fusion on features from different sources to generate a unified feature vector.

5. The method for predicting meniscus injury based on neural network according to claim 4, characterized in that: The construction of the meniscus injury prediction model based on the lightweight shallow neural network model also includes: using the BFGS optimization algorithm to update the network weights.

6. The method for predicting meniscus injury based on neural network according to claim 1, characterized in that: The meniscus injury prediction model outputs a current injury diagnosis and a future injury risk assessment.

7. The method for predicting meniscus injury based on neural network according to claim 1, characterized in that: Step S5 further includes performing a prediction accuracy assessment based on an AUC evaluation model, wherein the input content of the AUC evaluation model includes radiomic features, clinical features, and time dimension features; the radiomic features include radiomic features of the femur, femoral cartilage, tibia, tibial cartilage, medial meniscus, and lateral meniscus, namely, shape features, texture features, and intensity features; Clinical characteristics included age, gender, body mass index, history of trauma, meniscus injury, and knee surgery; temporal characteristics included baseline MRI imaging data and subsequent follow-up data.

8. A neural network-based meniscus injury prediction system, based on the neural network-based meniscus injury prediction method according to any one of claims 1 to 7, comprising: An image acquisition module, which is used to acquire knee MRI images; An image segmentation module for automatically segmenting the knee MRI image based on a convolutional neural network to identify and separate the weight-bearing structures of the knee joint; including automatically segmenting the MRI based on two-dimensional and three-dimensional convolutional neural networks, adjusting the segmentation results generated by the convolutional neural network based on a statistical shape model, and determining the segmentation accuracy; Radiomics feature extraction module, which is used to extract multiple image features based on the segmented image, including shape features, texture features, and image intensity features; An injury prediction model construction module is used to screen out features that are significantly correlated with meniscus injury based on the minimum absolute value convergence and selection operator algorithm; a meniscus injury prediction model is constructed based on a lightweight shallow neural network model; The prediction model training and validation module is used to train the model using the knee MRI image dataset and to validate and optimize the model using internal and external datasets.

9. A computer-readable storage medium, characterized in that The computer-readable storage medium stores program instructions of a neural network-based meniscus injury prediction method, and the program instructions of the neural network-based meniscus injury prediction method can be executed by one or more processors to implement the steps of the neural network-based meniscus injury prediction method as described in one of claims 1-7.

Images